Same kernel filter at the 1st epoch and the final epoch

[yananliusdu]

Hi there,

thanks for your implementation of binary connect.

I got a weird problem. I use your binary connect to train a model based on the MNIST. however, although I can get a model with an accuracy of over 90%. I find the kernel filters are quite random which are not apparently like edge detection. in addition, the kernel filter trained after 1st epoch and final epoch is almost identical, which means the convolutional kernel filters are not updated during training.

Do you know why that happens?

thanks.

[eghouti]

Can you share your code?

[yananliusdu]

Hi, according to the BinaryConnect paper, after training, the distribution of weight should be around -1 and 1 ( ). I tried to use your implementation of BinaryConnect on 10 letters from EMNIST. however, the weights are not like that in the paper.

do you know why? many thanks.